drivers/staging/lustre/lustre/ptlrpc/niobuf.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #define DEBUG_SUBSYSTEM S_RPC
  38 #include <obd_support.h>
  39 #include <lustre_net.h>
  40 #include <lustre_lib.h>
  41 #include <obd.h>
  42 #include <obd_class.h>
  43 #include "ptlrpc_internal.h"
  44
  45 /**
  46  * Helper function. Sends \a len bytes from \a base at offset \a offset
  47  * over \a conn connection to portal \a portal.
  48  * Returns 0 on success or error code.
  49  */
  50 static int ptl_send_buf (lnet_handle_md_t *mdh, void *base, int len,
  51                          lnet_ack_req_t ack, struct ptlrpc_cb_id *cbid,
  52                          struct ptlrpc_connection *conn, int portal, __u64 xid,
  53                          unsigned int offset)
  54 {
  55         int           rc;
  56         lnet_md_t        md;
  57
  58         LASSERT (portal != 0);
  59         LASSERT (conn != NULL);
  60         CDEBUG (D_INFO, "conn=%p id %s\n", conn, libcfs_id2str(conn->c_peer));
  61         md.start     = base;
  62         md.length    = len;
  63         md.threshold = (ack == LNET_ACK_REQ) ? 2 : 1;
  64         md.options   = PTLRPC_MD_OPTIONS;
  65         md.user_ptr  = cbid;
  66         md.eq_handle = ptlrpc_eq_h;
  67
  68         if (unlikely(ack == LNET_ACK_REQ &&
  69                      OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_ACK, OBD_FAIL_ONCE))){
  70                 /* don't ask for the ack to simulate failing client */
  71                 ack = LNET_NOACK_REQ;
  72         }
  73
  74         rc = LNetMDBind (md, LNET_UNLINK, mdh);
  75         if (unlikely(rc != 0)) {
  76                 CERROR ("LNetMDBind failed: %d\n", rc);
  77                 LASSERT (rc == -ENOMEM);
  78                 return -ENOMEM;
  79         }
  80
  81         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid "LPD64", offset %u\n",
  82                len, portal, xid, offset);
  83
  84         rc = LNetPut (conn->c_self, *mdh, ack,
  85                       conn->c_peer, portal, xid, offset, 0);
  86         if (unlikely(rc != 0)) {
  87                 int rc2;
  88                 /* We're going to get an UNLINK event when I unlink below,
  89                  * which will complete just like any other failed send, so
  90                  * I fall through and return success here! */
  91                 CERROR("LNetPut(%s, %d, "LPD64") failed: %d\n",
  92                        libcfs_id2str(conn->c_peer), portal, xid, rc);
  93                 rc2 = LNetMDUnlink(*mdh);
  94                 LASSERTF(rc2 == 0, "rc2 = %d\n", rc2);
  95         }
  96
  97         return 0;
  98 }
  99
 100 static void mdunlink_iterate_helper(lnet_handle_md_t *bd_mds, int count)
 101 {
 102         int i;
 103
 104         for (i = 0; i < count; i++)
 105                 LNetMDUnlink(bd_mds[i]);
 106 }
 107
 108
 109 /**
 110  * Register bulk at the sender for later transfer.
 111  * Returns 0 on success or error code.
 112  */
 113 int ptlrpc_register_bulk(struct ptlrpc_request *req)
 114 {
 115         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 116         lnet_process_id_t peer;
 117         int rc = 0;
 118         int rc2;
 119         int posted_md;
 120         int total_md;
 121         __u64 xid;
 122         lnet_handle_me_t  me_h;
 123         lnet_md_t        md;
 124
 125         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
 126                 return 0;
 127
 128         /* NB no locking required until desc is on the network */
 129         LASSERT(desc->bd_nob > 0);
 130         LASSERT(desc->bd_md_count == 0);
 131         LASSERT(desc->bd_md_max_brw <= PTLRPC_BULK_OPS_COUNT);
 132         LASSERT(desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
 133         LASSERT(desc->bd_req != NULL);
 134         LASSERT(desc->bd_type == BULK_PUT_SINK ||
 135                 desc->bd_type == BULK_GET_SOURCE);
 136
 137         /* cleanup the state of the bulk for it will be reused */
 138         if (req->rq_resend || req->rq_send_state == LUSTRE_IMP_REPLAY)
 139                 desc->bd_nob_transferred = 0;
 140         else
 141                 LASSERT(desc->bd_nob_transferred == 0);
 142
 143         desc->bd_failure = 0;
 144
 145         peer = desc->bd_import->imp_connection->c_peer;
 146
 147         LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback);
 148         LASSERT(desc->bd_cbid.cbid_arg == desc);
 149
 150         /* An XID is only used for a single request from the client.
 151          * For retried bulk transfers, a new XID will be allocated in
 152          * in ptlrpc_check_set() if it needs to be resent, so it is not
 153          * using the same RDMA match bits after an error.
 154          *
 155          * For multi-bulk RPCs, rq_xid is the last XID needed for bulks. The
 156          * first bulk XID is power-of-two aligned before rq_xid. LU-1431 */
 157         xid = req->rq_xid & ~((__u64)desc->bd_md_max_brw - 1);
 158         LASSERTF(!(desc->bd_registered &&
 159                    req->rq_send_state != LUSTRE_IMP_REPLAY) ||
 160                  xid != desc->bd_last_xid,
 161                  "registered: %d  rq_xid: "LPU64" bd_last_xid: "LPU64"\n",
 162                  desc->bd_registered, xid, desc->bd_last_xid);
 163
 164         total_md = (desc->bd_iov_count + LNET_MAX_IOV - 1) / LNET_MAX_IOV;
 165         desc->bd_registered = 1;
 166         desc->bd_last_xid = xid;
 167         desc->bd_md_count = total_md;
 168         md.user_ptr = &desc->bd_cbid;
 169         md.eq_handle = ptlrpc_eq_h;
 170         md.threshold = 1;                      /* PUT or GET */
 171
 172         for (posted_md = 0; posted_md < total_md; posted_md++, xid++) {
 173                 md.options = PTLRPC_MD_OPTIONS |
 174                              ((desc->bd_type == BULK_GET_SOURCE) ?
 175                               LNET_MD_OP_GET : LNET_MD_OP_PUT);
 176                 ptlrpc_fill_bulk_md(&md, desc, posted_md);
 177
 178                 rc = LNetMEAttach(desc->bd_portal, peer, xid, 0,
 179                                   LNET_UNLINK, LNET_INS_AFTER, &me_h);
 180                 if (rc != 0) {
 181                         CERROR("%s: LNetMEAttach failed x"LPU64"/%d: rc = %d\n",
 182                                desc->bd_export->exp_obd->obd_name, xid,
 183                                posted_md, rc);
 184                         break;
 185                 }
 186
 187                 /* About to let the network at it... */
 188                 rc = LNetMDAttach(me_h, md, LNET_UNLINK,
 189                                   &desc->bd_mds[posted_md]);
 190                 if (rc != 0) {
 191                         CERROR("%s: LNetMDAttach failed x"LPU64"/%d: rc = %d\n",
 192                                desc->bd_export->exp_obd->obd_name, xid,
 193                                posted_md, rc);
 194                         rc2 = LNetMEUnlink(me_h);
 195                         LASSERT(rc2 == 0);
 196                         break;
 197                 }
 198         }
 199
 200         if (rc != 0) {
 201                 LASSERT(rc == -ENOMEM);
 202                 spin_lock(&desc->bd_lock);
 203                 desc->bd_md_count -= total_md - posted_md;
 204                 spin_unlock(&desc->bd_lock);
 205                 LASSERT(desc->bd_md_count >= 0);
 206                 mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 207                 req->rq_status = -ENOMEM;
 208                 return -ENOMEM;
 209         }
 210
 211         /* Set rq_xid to matchbits of the final bulk so that server can
 212          * infer the number of bulks that were prepared */
 213         req->rq_xid = --xid;
 214         LASSERTF(desc->bd_last_xid == (req->rq_xid & PTLRPC_BULK_OPS_MASK),
 215                  "bd_last_xid = x"LPU64", rq_xid = x"LPU64"\n",
 216                  desc->bd_last_xid, req->rq_xid);
 217
 218         spin_lock(&desc->bd_lock);
 219         /* Holler if peer manages to touch buffers before he knows the xid */
 220         if (desc->bd_md_count != total_md)
 221                 CWARN("%s: Peer %s touched %d buffers while I registered\n",
 222                       desc->bd_export->exp_obd->obd_name, libcfs_id2str(peer),
 223                       total_md - desc->bd_md_count);
 224         spin_unlock(&desc->bd_lock);
 225
 226         CDEBUG(D_NET, "Setup %u bulk %s buffers: %u pages %u bytes, "
 227                "xid x"LPX64"-"LPX64", portal %u\n", desc->bd_md_count,
 228                desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink",
 229                desc->bd_iov_count, desc->bd_nob,
 230                desc->bd_last_xid, req->rq_xid, desc->bd_portal);
 231
 232         return 0;
 233 }
 234 EXPORT_SYMBOL(ptlrpc_register_bulk);
 235
 236 /**
 237  * Disconnect a bulk desc from the network. Idempotent. Not
 238  * thread-safe (i.e. only interlocks with completion callback).
 239  * Returns 1 on success or 0 if network unregistration failed for whatever
 240  * reason.
 241  */
 242 int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 243 {
 244         struct ptlrpc_bulk_desc *desc = req->rq_bulk;
 245         wait_queue_head_t            *wq;
 246         struct l_wait_info       lwi;
 247         int                   rc;
 248
 249         LASSERT(!in_interrupt());     /* might sleep */
 250
 251         /* Let's setup deadline for reply unlink. */
 252         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 253             async && req->rq_bulk_deadline == 0)
 254                 req->rq_bulk_deadline = cfs_time_current_sec() + LONG_UNLINK;
 255
 256         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 257                 return 1;                               /* never registered */
 258
 259         LASSERT(desc->bd_req == req);  /* bd_req NULL until registered */
 260
 261         /* the unlink ensures the callback happens ASAP and is the last
 262          * one.  If it fails, it must be because completion just happened,
 263          * but we must still l_wait_event() in this case to give liblustre
 264          * a chance to run client_bulk_callback() */
 265         mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
 266
 267         if (ptlrpc_client_bulk_active(req) == 0)        /* completed or */
 268                 return 1;                               /* never registered */
 269
 270         /* Move to "Unregistering" phase as bulk was not unlinked yet. */
 271         ptlrpc_rqphase_move(req, RQ_PHASE_UNREGISTERING);
 272
 273         /* Do not wait for unlink to finish. */
 274         if (async)
 275                 return 0;
 276
 277         if (req->rq_set != NULL)
 278                 wq = &req->rq_set->set_waitq;
 279         else
 280                 wq = &req->rq_reply_waitq;
 281
 282         for (;;) {
 283                 /* Network access will complete in finite time but the HUGE
 284                  * timeout lets us CWARN for visibility of sluggish NALs */
 285                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
 286                                            cfs_time_seconds(1), NULL, NULL);
 287                 rc = l_wait_event(*wq, !ptlrpc_client_bulk_active(req), &lwi);
 288                 if (rc == 0) {
 289                         ptlrpc_rqphase_move(req, req->rq_next_phase);
 290                         return 1;
 291                 }
 292
 293                 LASSERT(rc == -ETIMEDOUT);
 294                 DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
 295                           desc);
 296         }
 297         return 0;
 298 }
 299 EXPORT_SYMBOL(ptlrpc_unregister_bulk);
 300
 301 static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
 302 {
 303         struct ptlrpc_service_part      *svcpt = req->rq_rqbd->rqbd_svcpt;
 304         struct ptlrpc_service           *svc = svcpt->scp_service;
 305         int service_time = max_t(int, cfs_time_current_sec() -
 306                                  req->rq_arrival_time.tv_sec, 1);
 307
 308         if (!(flags & PTLRPC_REPLY_EARLY) &&
 309             (req->rq_type != PTL_RPC_MSG_ERR) &&
 310             (req->rq_reqmsg != NULL) &&
 311             !(lustre_msg_get_flags(req->rq_reqmsg) &
 312               (MSG_RESENT | MSG_REPLAY |
 313                MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
 314                 /* early replies, errors and recovery requests don't count
 315                  * toward our service time estimate */
 316                 int oldse = at_measured(&svcpt->scp_at_estimate, service_time);
 317
 318                 if (oldse != 0) {
 319                         DEBUG_REQ(D_ADAPTTO, req,
 320                                   "svc %s changed estimate from %d to %d",
 321                                   svc->srv_name, oldse,
 322                                   at_get(&svcpt->scp_at_estimate));
 323                 }
 324         }
 325         /* Report actual service time for client latency calc */
 326         lustre_msg_set_service_time(req->rq_repmsg, service_time);
 327         /* Report service time estimate for future client reqs, but report 0
 328          * (to be ignored by client) if it's a error reply during recovery.
 329          * (bz15815) */
 330         if (req->rq_type == PTL_RPC_MSG_ERR &&
 331             (req->rq_export == NULL || req->rq_export->exp_obd->obd_recovering))
 332                 lustre_msg_set_timeout(req->rq_repmsg, 0);
 333         else
 334                 lustre_msg_set_timeout(req->rq_repmsg,
 335                                        at_get(&svcpt->scp_at_estimate));
 336
 337         if (req->rq_reqmsg &&
 338             !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
 339                 CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
 340                        "req_flags=%#x magic=%d:%x/%x len=%d\n",
 341                        flags, lustre_msg_get_flags(req->rq_reqmsg),
 342                        lustre_msg_is_v1(req->rq_reqmsg),
 343                        lustre_msg_get_magic(req->rq_reqmsg),
 344                        lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
 345         }
 346 }
 347
 348 /**
 349  * Send request reply from request \a req reply buffer.
 350  * \a flags defines reply types
 351  * Returns 0 on sucess or error code
 352  */
 353 int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
 354 {
 355         struct ptlrpc_reply_state *rs = req->rq_reply_state;
 356         struct ptlrpc_connection  *conn;
 357         int                     rc;
 358
 359         /* We must already have a reply buffer (only ptlrpc_error() may be
 360          * called without one). The reply generated by sptlrpc layer (e.g.
 361          * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
 362          * have a request buffer which is either the actual (swabbed) incoming
 363          * request, or a saved copy if this is a req saved in
 364          * target_queue_final_reply().
 365          */
 366         LASSERT (req->rq_no_reply == 0);
 367         LASSERT (req->rq_reqbuf != NULL);
 368         LASSERT (rs != NULL);
 369         LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
 370         LASSERT (req->rq_repmsg != NULL);
 371         LASSERT (req->rq_repmsg == rs->rs_msg);
 372         LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
 373         LASSERT (rs->rs_cb_id.cbid_arg == rs);
 374
 375         /* There may be no rq_export during failover */
 376
 377         if (unlikely(req->rq_export && req->rq_export->exp_obd &&
 378                      req->rq_export->exp_obd->obd_fail)) {
 379                 /* Failed obd's only send ENODEV */
 380                 req->rq_type = PTL_RPC_MSG_ERR;
 381                 req->rq_status = -ENODEV;
 382                 CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
 383                        req->rq_export->exp_obd->obd_minor);
 384         }
 385
 386         /* In order to keep interoprability with the client (< 2.3) which
 387          * doesn't have pb_jobid in ptlrpc_body, We have to shrink the
 388          * ptlrpc_body in reply buffer to ptlrpc_body_v2, otherwise, the
 389          * reply buffer on client will be overflow.
 390          *
 391          * XXX Remove this whenver we drop the interoprability with such client.
 392          */
 393         req->rq_replen = lustre_shrink_msg(req->rq_repmsg, 0,
 394                                            sizeof(struct ptlrpc_body_v2), 1);
 395
 396         if (req->rq_type != PTL_RPC_MSG_ERR)
 397                 req->rq_type = PTL_RPC_MSG_REPLY;
 398
 399         lustre_msg_set_type(req->rq_repmsg, req->rq_type);
 400         lustre_msg_set_status(req->rq_repmsg,
 401                               ptlrpc_status_hton(req->rq_status));
 402         lustre_msg_set_opc(req->rq_repmsg,
 403                 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
 404
 405         target_pack_pool_reply(req);
 406
 407         ptlrpc_at_set_reply(req, flags);
 408
 409         if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
 410                 conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
 411         else
 412                 conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
 413
 414         if (unlikely(conn == NULL)) {
 415                 CERROR("not replying on NULL connection\n"); /* bug 9635 */
 416                 return -ENOTCONN;
 417         }
 418         ptlrpc_rs_addref(rs);              /* +1 ref for the network */
 419
 420         rc = sptlrpc_svc_wrap_reply(req);
 421         if (unlikely(rc))
 422                 goto out;
 423
 424         req->rq_sent = cfs_time_current_sec();
 425
 426         rc = ptl_send_buf (&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
 427                            (rs->rs_difficult && !rs->rs_no_ack) ?
 428                            LNET_ACK_REQ : LNET_NOACK_REQ,
 429                            &rs->rs_cb_id, conn,
 430                            ptlrpc_req2svc(req)->srv_rep_portal,
 431                            req->rq_xid, req->rq_reply_off);
 432 out:
 433         if (unlikely(rc != 0))
 434                 ptlrpc_req_drop_rs(req);
 435         ptlrpc_connection_put(conn);
 436         return rc;
 437 }
 438 EXPORT_SYMBOL(ptlrpc_send_reply);
 439
 440 int ptlrpc_reply (struct ptlrpc_request *req)
 441 {
 442         if (req->rq_no_reply)
 443                 return 0;
 444         else
 445                 return (ptlrpc_send_reply(req, 0));
 446 }
 447 EXPORT_SYMBOL(ptlrpc_reply);
 448
 449 /**
 450  * For request \a req send an error reply back. Create empty
 451  * reply buffers if necessary.
 452  */
 453 int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
 454 {
 455         int rc;
 456
 457         if (req->rq_no_reply)
 458                 return 0;
 459
 460         if (!req->rq_repmsg) {
 461                 rc = lustre_pack_reply(req, 1, NULL, NULL);
 462                 if (rc)
 463                         return rc;
 464         }
 465
 466         if (req->rq_status != -ENOSPC && req->rq_status != -EACCES &&
 467             req->rq_status != -EPERM && req->rq_status != -ENOENT &&
 468             req->rq_status != -EINPROGRESS && req->rq_status != -EDQUOT)
 469                 req->rq_type = PTL_RPC_MSG_ERR;
 470
 471         rc = ptlrpc_send_reply(req, may_be_difficult);
 472         return rc;
 473 }
 474 EXPORT_SYMBOL(ptlrpc_send_error);
 475
 476 int ptlrpc_error(struct ptlrpc_request *req)
 477 {
 478         return ptlrpc_send_error(req, 0);
 479 }
 480 EXPORT_SYMBOL(ptlrpc_error);
 481
 482 /**
 483  * Send request \a request.
 484  * if \a noreply is set, don't expect any reply back and don't set up
 485  * reply buffers.
 486  * Returns 0 on success or error code.
 487  */
 488 int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
 489 {
 490         int rc;
 491         int rc2;
 492         int mpflag = 0;
 493         struct ptlrpc_connection *connection;
 494         lnet_handle_me_t  reply_me_h;
 495         lnet_md_t        reply_md;
 496         struct obd_device *obd = request->rq_import->imp_obd;
 497
 498         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
 499                 return 0;
 500
 501         LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
 502         LASSERT(request->rq_wait_ctx == 0);
 503
 504         /* If this is a re-transmit, we're required to have disengaged
 505          * cleanly from the previous attempt */
 506         LASSERT(!request->rq_receiving_reply);
 507
 508         if (request->rq_import->imp_obd &&
 509             request->rq_import->imp_obd->obd_fail) {
 510                 CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
 511                        request->rq_import->imp_obd->obd_name);
 512                 /* this prevents us from waiting in ptlrpc_queue_wait */
 513                 request->rq_err = 1;
 514                 request->rq_status = -ENODEV;
 515                 return -ENODEV;
 516         }
 517
 518         connection = request->rq_import->imp_connection;
 519
 520         lustre_msg_set_handle(request->rq_reqmsg,
 521                               &request->rq_import->imp_remote_handle);
 522         lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
 523         lustre_msg_set_conn_cnt(request->rq_reqmsg,
 524                                 request->rq_import->imp_conn_cnt);
 525         lustre_msghdr_set_flags(request->rq_reqmsg,
 526                                 request->rq_import->imp_msghdr_flags);
 527
 528         if (request->rq_resend)
 529                 lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
 530
 531         if (request->rq_memalloc)
 532                 mpflag = cfs_memory_pressure_get_and_set();
 533
 534         rc = sptlrpc_cli_wrap_request(request);
 535         if (rc)
 536                 GOTO(out, rc);
 537
 538         /* bulk register should be done after wrap_request() */
 539         if (request->rq_bulk != NULL) {
 540                 rc = ptlrpc_register_bulk (request);
 541                 if (rc != 0)
 542                         GOTO(out, rc);
 543         }
 544
 545         if (!noreply) {
 546                 LASSERT (request->rq_replen != 0);
 547                 if (request->rq_repbuf == NULL) {
 548                         LASSERT(request->rq_repdata == NULL);
 549                         LASSERT(request->rq_repmsg == NULL);
 550                         rc = sptlrpc_cli_alloc_repbuf(request,
 551                                                       request->rq_replen);
 552                         if (rc) {
 553                                 /* this prevents us from looping in
 554                                  * ptlrpc_queue_wait */
 555                                 request->rq_err = 1;
 556                                 request->rq_status = rc;
 557                                 GOTO(cleanup_bulk, rc);
 558                         }
 559                 } else {
 560                         request->rq_repdata = NULL;
 561                         request->rq_repmsg = NULL;
 562                 }
 563
 564                 rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/
 565                                   connection->c_peer, request->rq_xid, 0,
 566                                   LNET_UNLINK, LNET_INS_AFTER, &reply_me_h);
 567                 if (rc != 0) {
 568                         CERROR("LNetMEAttach failed: %d\n", rc);
 569                         LASSERT (rc == -ENOMEM);
 570                         GOTO(cleanup_bulk, rc = -ENOMEM);
 571                 }
 572         }
 573
 574         spin_lock(&request->rq_lock);
 575         /* If the MD attach succeeds, there _will_ be a reply_in callback */
 576         request->rq_receiving_reply = !noreply;
 577         /* We are responsible for unlinking the reply buffer */
 578         request->rq_must_unlink = !noreply;
 579         /* Clear any flags that may be present from previous sends. */
 580         request->rq_replied = 0;
 581         request->rq_err = 0;
 582         request->rq_timedout = 0;
 583         request->rq_net_err = 0;
 584         request->rq_resend = 0;
 585         request->rq_restart = 0;
 586         request->rq_reply_truncate = 0;
 587         spin_unlock(&request->rq_lock);
 588
 589         if (!noreply) {
 590                 reply_md.start     = request->rq_repbuf;
 591                 reply_md.length    = request->rq_repbuf_len;
 592                 /* Allow multiple early replies */
 593                 reply_md.threshold = LNET_MD_THRESH_INF;
 594                 /* Manage remote for early replies */
 595                 reply_md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
 596                         LNET_MD_MANAGE_REMOTE |
 597                         LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
 598                 reply_md.user_ptr  = &request->rq_reply_cbid;
 599                 reply_md.eq_handle = ptlrpc_eq_h;
 600
 601                 /* We must see the unlink callback to unset rq_must_unlink,
 602                    so we can't auto-unlink */
 603                 rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
 604                                   &request->rq_reply_md_h);
 605                 if (rc != 0) {
 606                         CERROR("LNetMDAttach failed: %d\n", rc);
 607                         LASSERT (rc == -ENOMEM);
 608                         spin_lock(&request->rq_lock);
 609                         /* ...but the MD attach didn't succeed... */
 610                         request->rq_receiving_reply = 0;
 611                         spin_unlock(&request->rq_lock);
 612                         GOTO(cleanup_me, rc = -ENOMEM);
 613                 }
 614
 615                 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64
 616                        ", portal %u\n",
 617                        request->rq_repbuf_len, request->rq_xid,
 618                        request->rq_reply_portal);
 619         }
 620
 621         /* add references on request for request_out_callback */
 622         ptlrpc_request_addref(request);
 623         if (obd->obd_svc_stats != NULL)
 624                 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
 625                         atomic_read(&request->rq_import->imp_inflight));
 626
 627         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 628
 629         do_gettimeofday(&request->rq_arrival_time);
 630         request->rq_sent = cfs_time_current_sec();
 631         /* We give the server rq_timeout secs to process the req, and
 632            add the network latency for our local timeout. */
 633         request->rq_deadline = request->rq_sent + request->rq_timeout +
 634                 ptlrpc_at_get_net_latency(request);
 635
 636         ptlrpc_pinger_sending_on_import(request->rq_import);
 637
 638         DEBUG_REQ(D_INFO, request, "send flg=%x",
 639                   lustre_msg_get_flags(request->rq_reqmsg));
 640         rc = ptl_send_buf(&request->rq_req_md_h,
 641                           request->rq_reqbuf, request->rq_reqdata_len,
 642                           LNET_NOACK_REQ, &request->rq_req_cbid,
 643                           connection,
 644                           request->rq_request_portal,
 645                           request->rq_xid, 0);
 646         if (rc == 0)
 647                 GOTO(out, rc);
 648
 649         ptlrpc_req_finished(request);
 650         if (noreply)
 651                 GOTO(out, rc);
 652
 653  cleanup_me:
 654         /* MEUnlink is safe; the PUT didn't even get off the ground, and
 655          * nobody apart from the PUT's target has the right nid+XID to
 656          * access the reply buffer. */
 657         rc2 = LNetMEUnlink(reply_me_h);
 658         LASSERT (rc2 == 0);
 659         /* UNLINKED callback called synchronously */
 660         LASSERT(!request->rq_receiving_reply);
 661
 662  cleanup_bulk:
 663         /* We do sync unlink here as there was no real transfer here so
 664          * the chance to have long unlink to sluggish net is smaller here. */
 665         ptlrpc_unregister_bulk(request, 0);
 666  out:
 667         if (request->rq_memalloc)
 668                 cfs_memory_pressure_restore(mpflag);
 669         return rc;
 670 }
 671 EXPORT_SYMBOL(ptl_send_rpc);
 672
 673 /**
 674  * Register request buffer descriptor for request receiving.
 675  */
 676 int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
 677 {
 678         struct ptlrpc_service     *service = rqbd->rqbd_svcpt->scp_service;
 679         static lnet_process_id_t  match_id = {LNET_NID_ANY, LNET_PID_ANY};
 680         int                       rc;
 681         lnet_md_t                md;
 682         lnet_handle_me_t          me_h;
 683
 684         CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
 685                service->srv_req_portal);
 686
 687         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
 688                 return (-ENOMEM);
 689
 690         /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
 691          * which means buffer can only be attached on local CPT, and LND
 692          * threads can find it by grabbing a local lock */
 693         rc = LNetMEAttach(service->srv_req_portal,
 694                           match_id, 0, ~0, LNET_UNLINK,
 695                           rqbd->rqbd_svcpt->scp_cpt >= 0 ?
 696                           LNET_INS_LOCAL : LNET_INS_AFTER, &me_h);
 697         if (rc != 0) {
 698                 CERROR("LNetMEAttach failed: %d\n", rc);
 699                 return (-ENOMEM);
 700         }
 701
 702         LASSERT(rqbd->rqbd_refcount == 0);
 703         rqbd->rqbd_refcount = 1;
 704
 705         md.start     = rqbd->rqbd_buffer;
 706         md.length    = service->srv_buf_size;
 707         md.max_size  = service->srv_max_req_size;
 708         md.threshold = LNET_MD_THRESH_INF;
 709         md.options   = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
 710         md.user_ptr  = &rqbd->rqbd_cbid;
 711         md.eq_handle = ptlrpc_eq_h;
 712
 713         rc = LNetMDAttach(me_h, md, LNET_UNLINK, &rqbd->rqbd_md_h);
 714         if (rc == 0)
 715                 return (0);
 716
 717         CERROR("LNetMDAttach failed: %d; \n", rc);
 718         LASSERT (rc == -ENOMEM);
 719         rc = LNetMEUnlink (me_h);
 720         LASSERT (rc == 0);
 721         rqbd->rqbd_refcount = 0;
 722
 723         return (-ENOMEM);
 724 }